Probabilistic forecasting of cross-sectional returns: A Bayesian dynamic factor model with heteroskedasticity

Abstract

The M6 Financial Forecasting Competition forecasting track required probabilistic forecasting of monthly returns for a universe of 100 assets. This paper describes a Bayesian dynamic factor model with heteroskedasticity that was used to win the year-long forecasting track. The model’s strengths include modularity, handling of missing data, and regularization through hierarchical distributions. Probability modeling and recent advances in probabilistic programming languages make defining such models and performing inference straightforward.